Lumai, an optical compute company spun out of the University of Oxford in 2021, is moving its Iris Nova server out of the research setting. According to Data Center Knowledge, a cluster of the machines will be installed in the Scaling Inference Lab, a UK facility that evaluates new AI hardware on real workloads in a live commercial environment. The UK's Advanced Research and Invention Agency supports the lab, and the non-profit CommonAI CIC operates it. Iris Nova will be the first optical inference platform deployed there.
The design rests on a split in how inference works. Prefill, where the prompt and context are processed, is compute-bound, so the speed of the arithmetic sets overall performance. Decode, where output tokens are generated one at a time, is limited by memory bandwidth, which means faster math helps less. Lumai's Phillip Burr says Iris Nova is built for prefill only, with GPUs or other decode-optimized hardware handling the rest. Training is also a poor fit today, he said, because it needs higher precision and continuous weight updates.
The company says Iris encodes numbers as light intensity and uses passive optical parts such as lenses to perform matrix multiplication. Lumai claims this computation consumes negligible energy, with power rising only slightly as performance scales. The architecture is hybrid, with memory, control and nonlinear functions still handled electronically.
Operators would need disaggregated serving to route prefill and decode to separate hardware. Burr says Iris works with PyTorch, standard networking interfaces, and serving frameworks including vLLM and SGLang. He also says the servers use components already made at high volume for data center communications, and run in standard air-cooled racks without liquid cooling.
Burr argues the payoff depends on how much of an operator's compute goes to prefill, which he says dominates with long prompts and agentic workloads that repeatedly process large contexts. Moving prefill off GPUs would also free that GPU capacity for decode.
Lumai describes the deployment as a way to validate the system and gather system-level performance data to inform its next, production-oriented product.
Why it matters: Power budgets are fixed while inference demand grows, so any hardware that offloads work from GPUs deserves scrutiny. The efficiency claims so far come from the company, though, and the lab trial is meant to produce independent system-level data. Operators should watch those results, and how much of their own workload is prefill, before drawing conclusions.
TwinStack automated analysis, based on reporting by Data Center Knowledge. Check the original for full details.